Augustin
13.11 Augustin — Independent Judge¶
Responsibilities¶
- consume blinded Urabrask
QualityReportrecords; - apply hard eligibility requirements;
- apply the tail-risk veto before any utility comparison;
- enforce Tamiyo’s active strategic envelope;
- calculate adjudicated utility, risk and uncertainty charges;
- compare surviving candidates against mandatory no-op;
- apply independent admission-audit rules;
- return
ADMIT,NO_OP,REJECT,DEFERorREQUIRE_RETEST; - issue admission warrants;
- adjudicate continued tenancy of committed structures;
- issue maintenance warrants;
- calibrate thresholds on validation trajectories only;
- and emit explicit decision reasons and policy versions.
Lexicographic admission order (ADR-0004)¶
Admission is a risk judgement before it is a quality judgement. Augustin adjudicates in three ordered stages, and no later stage can reopen an earlier one:
- Hard eligibility. Conformance facts from the blinded
QualityReport: identity, compilation, runtime and gradient checks, determinism, evidence completeness for the assurance class, budget. - Tail-risk veto. Augustin estimates, per candidate, the tail of the intervention-outcome distribution — the catastrophic case, not the expectation — from certified evidence (integration shock, instability, numerical events, trajectory behaviour). A candidate whose tail estimate breaches the veto threshold is out of contention, and no measured benefit can offset the veto (INV-45). The assurance class owns the veto operating point, and the threshold is priced against current snapshot distance: the further the trajectory sits from its last snapshot, the more a rollback costs, the tighter the veto.
- Utility competition.
u_admitranks the survivors against mandatory no-op.
No-op never faces the veto — no intervention carries no intervention tail
risk — so it survives every stage by construction. A pool whose every
candidate is vetoed resolves to NO_OP (or DEFER/REQUIRE_RETEST under
policy), never to a least-bad survivor.
This order is what justifies the Momir/Augustin split: Momir optimises expected value and can afford to be wrong often, because its errors cost compute; Augustin bounds worst case and must be conservative, because its errors cost the host. They optimise different functionals of the same distribution.
Admission utility¶
For a candidate \(c\) that survived eligibility and the tail-risk veto, state \(s\), and horizon \(H\):
where:
- \(C_c\) is compute and latency cost;
- \(S_c\) is integration shock;
- \(P_c\) is parameter and memory cost;
- \(R_c\) is expected policy risk — tail risk is expressible only through the stage-2 veto, never as a utility charge (ADR-0004);
- \(U_c\) is uncertainty.
The no-op candidate has policy utility exactly zero.
Continued-tenancy utility¶
For a resident growth:
Installation shock is omitted because a resident growth is no longer integrating. Shared cost terms use shared weights unless a difference is structurally justified and pre-registered — the asymmetry between admission and retention lives in the thresholds, never in per-side weights.
Retention hysteresis (ADR-0005)¶
Admission and retention thresholds form a Schmitt trigger: admission requires the conservative margin over no-op to exceed \(\theta_{\mathrm{admit}}\); continued tenancy requires only \(u_{\mathrm{retain}} \ge \theta_{\mathrm{retain}}\), with \(\theta_{\mathrm{retain}} = \theta_{\mathrm{admit}} - \Delta\) and \(\Delta > 0\) strictly (INV-33). A resident growth that drifts modestly below the admission bar is not thereby lysed.
The hysteresis band \(\Delta\) is an explicit, versioned Augustin policy parameter, carried by adjudication_policy_version, and its width is measured against observed execution noise \(\sigma_{\mathrm{exec}}\) — sized so noise-driven estimate movement cannot cross both thresholds — never picked by feel. Cooldowns are a frequency limiter for pathological cases, not the stability mechanism: they are a time-domain patch and do not remove a threshold-domain instability.
Invariants¶
- Augustin does not execute tests, alter data, call kernels or rerun a branch.
- It cannot see candidate source during adjudication.
- It can select no-op even when Tamiyo allocated budget and Narset commissioned growth.
- Hard Urabrask defects make a candidate ineligible according to the applicable Leyline policy.
- The tail-risk veto is adjudicated before any utility comparison and cannot be traded against measured benefit (INV-45).
- Every veto, and every thin-margin pass, is recorded per candidate in
tail_veto_results(INV-31). - Decision thresholds are frozen before confirmatory runs.
- Every warrant is bound to a specific evidence digest, semantic hash, envelope and policy version.
Smell¶
If Augustin asks for a more favourable minibatch after seeing the evidence, the judge has tampered with the case.
If a candidate's measured benefit is cited as a reason to soften the veto, the judge has repriced catastrophe.